Models
Variational AutoencoderVAE
Also known as: variational autoencoder
A generative model that learns to compress data into a smooth "meaning space" and generate new samples from it; also a building block inside modern image generators.
A VAE is a type of generative model that learns to squeeze data (like images) down into a compact, organised "map", then generate new examples by picking points on that map. Because the map is smooth, nearby points give similar results, so you can blend and explore. On its own it is less famous than other generators, but a VAE-like component quietly powers many of today's image tools by handling that compression step.
A variational autoencoder (VAE) has two parts: - Encoder: compresses input into a compact latent representation (a smooth "meaning space"). - Decoder: reconstructs data from a point in that space.
The "variational" trick makes the latent space continuous and well-organised, so you can sample new points and decode them into new, plausible data, or interpolate smoothly between examples.
Where you meet them
VAEs were an early generative approach alongside GANs. Today their biggest role is often as the compression component inside latent diffusion models: the VAE maps images to a compact latent space where diffusion runs efficiently, then decodes back to pixels.
A VAE is a latent-variable generative model trained by maximising the evidence lower bound (ELBO), balancing reconstruction against a KL term that regularises the approximate posterior toward a prior, yielding a continuous, structured latent space; the reparameterisation trick enables gradient-based training (Kingma & Welling 2013). This gives principled likelihood-based training and smooth, interpolable latents, at the cost of characteristically blurrier samples than GANs due to the reconstruction/KL trade-off and posterior approximations.
While pure VAEs were surpassed for high-fidelity synthesis, VAE-style autoencoders are foundational to latent diffusion: a (often VQ-)autoencoder compresses images into a latent space where diffusion is far cheaper, decoding to pixels afterward, one of the key efficiency enablers of modern text-to-image systems. Variants (VQ-VAE, β-VAE) address discretisation and disentanglement. Conceptually, VAEs connect generative modelling, representation learning and variational inference.